AMD has unveiled Helios, its first rack-scale AI computing platform, positioning it as a direct challenge to Nvidia’s data-centre systems. The company says the 72-GPU system exceeds Nvidia’s Blackwell and forthcoming Vera Rubin platforms on several headline measures, including memory, bandwidth and training performance, although these comparisons remain vendor claims rather than independently verified results.

Helios uses AMD’s new Instinct MI455X accelerator, based on its fifth-generation CDNA architecture and built from 24 chiplets, including eight 2nm compute dies and 12 HBM4 memory stacks. AMD claims Helios offers 50% more HBM4 memory and scale-out bandwidth than Vera Rubin, 15–25% higher AI-training performance, and a 30% performance-per-dollar advantage; Nvidia is expected to retain a claimed 25% FP4 inference lead where its compression technology applies. The unusually large rack is 1.2 metres wide and 44 rack units high, nearly twice the size of Nvidia’s NVL72.